AI Agent Logging: Understand Actions Without Leaking Secrets
Design AI agent logs and run receipts that explain what ran, what changed and why it stopped, while keeping secrets and personal data out. Schema, redaction and retention.
The AGNT Journal · Engineering
Engineering
Field notes from building AGNT — decisions, trade-offs and lessons from production.
Engineering
Engineering
AI Agent Permissions: Read, Draft, Approve, Execute
Design AI agent permissions enforced in code, not prompts: action classes, exact-action approval, expiry, one-time use, revocation and the negative tests that prove it.
Engineering
Why Your AI Agent Loops, Forgets or Claims It Is Finished
Diagnose AI agent failures by symptom: loops, forgetting, stale facts, false completion, duplicate actions and runaway cost. Find the responsible layer, fix it and add a regression test.
Engineering
AI Harness Engineering: How to Build AI Agents Your Business Can Trust (2027 Guide)
AI harness engineering for teams: the permissions, memory, evals and cost limits that make AI agents safe on business data. Sourced guide with a tested starter.
Engineering
What Does a Personal AI Assistant Cost? A Reproducible Worksheet
Work out what a personal AI assistant really costs per successful task: tokens, tools, hosting, hardware, energy and your time, with formulas, scenarios and budget limits.
Engineering
How to Test a Personal AI Assistant Before Trusting It
Evaluate a personal AI assistant on outcomes, not fluency: a 30-case suite, final-state checks, the right graders, pass@k vs pass^k, honest reporting and regression tests.
Engineering
Personal AI Memory: Store, Inspect, Correct, Export and Delete
Design personal AI memory you can see and change: what to remember, a record format with provenance, correction you can verify, portable export and honest deletion limits.
Engineering
Prompt Injection in Personal Assistants: A Safe Testing Lab
A defensive lab for prompt injection in personal AI assistants: synthetic fixtures, layered controls, scoring security and utility separately, and what passing does not prove.
Engineering
Scheduling AI Workflows Without Duplicate Actions
Run scheduled AI workflows that never act twice: separate schedules from runs, checkpoint state, use idempotency keys, classify retries, reconcile uncertain outcomes and test failures.
Engineering
AI Agent Observability: See Every Trace, Tool Call and Cost
AI agent observability explained: what to trace, the OpenTelemetry GenAI conventions, the metrics that matter, evals vs tracing, privacy, and how teams debug agents in production.
Engineering
AI Agents vs Workflows: When to Use Each (With Examples)
AI agent vs workflow, explained with Anthropic's definitions: when a fixed workflow beats an agent, when an agent is worth the cost, the five workflow patterns, and how to combine both.
Engineering
Single-Agent vs Multi-Agent Systems: When One Agent Is Enough
Single-agent vs multi-agent AI systems compared with evidence from Anthropic, Cognition and OpenAI: when multiple agents win, what they cost in tokens, and the patterns that actually work.
Engineering
Model vs. Harness: Where Super Intelligence Actually Does Work
The model thinks. The harness works. A plain-language breakdown of what a language model does, what an agent harness does, how one real task splits between them, and why the harness decides whether super intelligence ships results or just talks.
Engineering
The NN-Shaped Hammer: When to Learn and When to Just Write the Damn Code
Every learned parameter is a liability. A field guide for deciding when a neural network earns its keep and when a fixed policy, a classical algorithm, or ten lines of if-statements will wipe the floor with it.
Engineering
Pi-AI Provider Migration: Decision Report — Why We're Not Migrating
A thorough technical evaluation of migrating AGNT's LLM provider system to @mariozechner/pi-ai. After quantifying maintenance costs, migration effort, ROI, and dependency risks, we recommend against migration and propose targeted improvements instead.
No articles in this category yet.